ERNIE 5.1
Baidu · 2026-05-09
ERNIE 5.1 is Baidu's text-to-text MoE foundation model, derived by extracting an optimal subnetwork from ERNIE 5.0's elastic model matrix and served as the closed, hosted `ernie-5.1` API model. Baidu reports that its total and activated parameter counts are approximately one-third and one-half of ERNIE 5.0's, respectively; its post-training combines SFT, domain-expert training, on-policy distillation, and online reinforcement learning, while Qianfan exposes chain-of-thought output with a configurable reasoning-token budget.